- Email: [email protected]
The potential of Raman spectroscopy for characterisation of the fatty acid unsaturation of salmon
Analytica Chimica Acta 572 (2006) 85–92
The potential of Raman spectroscopy for characterisation of the fatty acid unsaturation of salmon Nils Kristian Afseth a,b,∗ , Jens Petter Wold a , Vegard Herman Segtnan a a
b
˚ Norway MATFORSK, Norwegian Food Research Institute, Osloveien 1, N-1430 As, ˚ Norway Department of Chemistry, Biotechnology and Food Science, Norwegian University of Life Sciences, As, Received 8 February 2006; received in revised form 3 May 2006; accepted 4 May 2006 Available online 10 May 2006
Abstract Raman spectroscopy has been evaluated for characterisation of the degree of fatty acid unsaturation (iodine value) of salmon (Salmo salar). The Norwegian Quality Cuts from 50 salmon samples were obtained, and the samples provided an iodine value range of 147.8–170.0 g I2 /100 g fat, reflecting a normal variation of farmed salmon. Raman measurements were performed both on different spots of the intact salmon muscle, on ground salmon samples as well as on oil extracts, and partial least squares regression (PLSR) was utilised for calibration. The oil spectra provided better iodine value predictions than the other data sets, and a correlation coefficient of 0.87 with a root mean square error of cross-validation of 2.5 g I2 /100 g fat was achieved using only one PLSR component. The ground samples provided comparable results, but at least two PLSR components were needed. Higher prediction errors were obtained from Raman spectra of intact salmon muscle, and this may partly be explained by sampling uncertainties in the relation between Raman measurements and reference analysis. All PLSR models obtained were based on chemically sound regression coefficients, and thus information regarding fatty acid unsaturation is readily available from Raman spectra even in systems with high contents of protein and water. The accuracy, the robustness and the low complexity of the PLSR models obtained suggest Raman spectroscopy as a promising method for rapid in-process control of the degree of unsaturation in salmon samples. © 2006 Elsevier B.V. All rights reserved. Keywords: Raman spectroscopy; Salmon; Iodine value; Partial least squares regression; Sampling; In-process control
1. Introduction There is an increasing interest in potential methods for fast and non-destructive characterisation of the fatty acid composition of marine food products. This is related to the great commercial value connected to dietary fats of marine origin in general, and the presence of several essential health promoting fatty acids like eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA) in particular. If reliable in-process techniques for control of fatty acid features are implemented, this will benefit both producers and consumers. For the producers, in-process control may provide better means for sorting raw materials as well as provide better documentation of product quality for price differentiation and promotion. In addition, rapid control of fatty acid composition is essential when controlling the effect of different feeding regimes. Ever increasing consumer demands for
∗
Corresponding author. Tel.: +47 64 97 04 18; fax: +47 64 97 03 33. E-mail address: [email protected] (N.K. Afseth).
0003-2670/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2006.05.013
improved food component labelling will also be met. At present, gas chromatography (GC) is the method of choice for analysing the fatty acid composition of foods. This method, however, is destructive and involves laborious wet chemistry for sample preparation. The results obtained are accurate, but the procedure is often too costly and time consuming for industrial quality evaluation. For several years Raman spectroscopy has been referred to as a promising method for rapid and non-destructive analysis of foods [1,2]. The technique is based on measuring fundamental vibrational modes of chemical bonds, and quantitative as well as qualitative structural and chemical information of a given sample may be obtained. What makes Raman spectroscopy especially suitable for food analysis is that minimal or no sample preparation is required, and that measurements may be carried out utilising fibre optics. In addition, the influence of water on the Raman spectra is almost negligible. Interference from strong and broad fluorescence signals combined with low sensitivity have traditionally been some of the major challenges. However, instrumental improvements concerning
86
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
both sampling techniques, stability and sensitivity have been achieved, paving the way for an increasing number of articles related to Raman spectroscopy and food analysis in recent years [3,4]. A range of feasibility studies has shown the potential of employing Raman spectroscopy as a means for quantifying different parameters related to fat composition. In medical diagnostics, quantification of different lipid groups like phospholipids, cholesterol, cholesterolesters and various fatty acids in arteries, skin layers or other kinds of healthy and diseased tissues has been accomplished [5–7]. In food science the focus has mainly been on the analysis of pure oils and fats. First and foremost, bulk parameters like the iodine value have been shown to correlate well with Raman spectra [8,9]. This is mainly related to the favourable Raman cross sections of chemical bands originating from the unsaturation of carbon chains. Good quantitative descriptions of the cis/trans-isomer ratio of fats and oils [10,11] as well as the level of conjugated fatty acids in milk fat [12] have also been obtained. In addition, oil and fat classification [13] and predictions of various single fatty acids have been carried out [14]. However, there are few studies published focusing on Raman analysis of fat composition of fish and animal adipose tissue, and to the knowledge of the authors there are no recorded references utilising Raman spectroscopy for quantifying fat composition of food systems with high contents of water and proteins. Traditionally, the iodine value has been one of the standard parameters related to chemical composition and quality of fats and oils. The analytical test is based on the ability of elemental iodine (I2 ) to quantitatively add to carbon–carbon double bonds, and the parameter is defined as the weight of iodine absorbed in an amount of 100 g of fat sample. Thus, as long as there are no other compounds present containing carbon–carbon double bonds, the iodine value is a direct measure of the level of fatty acid unsaturation [15]. Spectroscopic techniques like near-infrared (NIR) and mid-infrared (IR) spectroscopy both have been used for rapid determination of the iodine value of fats and oils with high accuracy [16,17]. However, strong water absorbances severely influence both NIR and IR spectra. In addition, protein absorbances often may obscure NIR bands related to fatty acid unsaturation [18]. Thus, these techniques have major disadvantages when analysing samples with high contents of water and proteins. Similar disturbances may to a great extent be avoided using Raman spectroscopy. In a previous study, the authors showed that Raman spectroscopy may be used to quantify bulk fatty acid features like the iodine value and the amount of saturated, monounsaturated and polyunsaturated fatty acids of complex food model samples with high contents of proteins and water [18]. The samples had a wide range in the variation of the fatty acid composition, and important spectral characteristics regarding the fatty acid parameters were obtained. In the present study a set of salmon samples with naturally varying fatty acid composition and overall sample heterogeneity has been analysed. The main objectives of the study were two-fold. First, the performance of Raman spectroscopy for modelling and prediction of the iodine
values of salmon was investigated. Second, by retrieving spectra from different parts of the intact salmon muscle, ground samples and oil extracts, a comparison of different sampling regimes for spectral acquisition was conducted. 2. Experimental 2.1. Materials Fifty samples of Atlantic salmon (S. salar) were obtained from the Nutreco Aquaculture Research Centre (Stavanger, Norway). The salmon samples originated from 13 different farms, and they were collected during a period of approximately 2 months. The samples were selected in order to span the range of the fatty acid composition normally occurring in farmed salmon. Muscle samples from the region of the Norwegian Quality Cut (NQC) were obtained [19], and the NQCs were split in half. One half of the NQC was left for reference analysis at Nutreco, whereas the other half was shipped frozen in enclosed plastic ˚ Norway) for Raman spectroscopic analbags to Matforsk (As, ysis. The location and the splitting of the NQCs are indicated in Fig. 1. 2.2. Reference analysis The fatty acid composition of the salmon samples were determined using a standard gas chromatography (GC) procedure. Prior to sample homogenisation the skin of the NQC was removed such that neither skin remnants were left at the muscle nor muscle remnants were left at the skin. Fat extraction
Fig. 1. Location and anatomy of the Norwegian Quality Cut (NQC) of salmon. The definitions used in the current study are provided in parenthesis. The dotted square indicates the division of NQCs into samples for reference and spectroscopic measurements, respectively.
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
and conversion to methyl esters were performed as described in the literature [20]. Methanolic hydrochloric acid (2.15 mL, 3N) was added to a sample of homogenised salmon (200 mg) and left enclosed over night. Water (1 mL) was added, and the methyl esters were extracted from the methanol phase with hexane (2 mL). The sample was centrifuged twice, with collection of the hexane phase and re-addition of hexane (2 mL) in between. The final hexane phase (25 L) was diluted with isooctane (1 mL) and injected on a Perkin-Elmer Autosystem GC equipped with a PSS injector, a CP-Wax 52 CB column (L = 25 m; i.d. = 0.25 mm; d.f. = 0.20 m, Varian), a flame ionisation detector and He as carrier gas. Extraction and GC-analysis were performed in duplicates of all samples. The compounds were identified and system performance was checked using standard samples. The concentration of individual fatty acids was expressed in percentage of total fat content. The iodine value was calculated directly from the GC results by converting the concentration of 38 fatty acids (expressed in percentage of total fat content) to the corresponding number of carbon–carbon double bonds (expressed in mol). The conversion was effected employing the molar mass of each corresponding fatty acid methyl ester. The iodine value then was calculated as the weight of elemental iodine corresponding to the total number of carbon–carbon double bonds. The uncertainty of the calculated iodine value was estimated using the Gaussian law of error propagation [21], whereas uncertainty estimates of each individual fatty acid were calculated on the basis of replicate measurements of a routine fatty acid standard. 2.3. Sample preparation A total of five data sets of Raman spectra were obtained from the salmon samples: one data set from extracted oils; one data set from ground salmon samples and three different data sets from intact samples. The main parts of the salmon NQCs are shown in Fig. 1. Lipids are mainly located in the dorsal fat depot, in the belly flap, in the collagenous connective tissue (myocommata) throughout the light and dark muscles, as well as in the intermediate space of myofibrils in the dark muscle [22]. The salmon samples showed great variability in both size and structure. Thus, three main regions were identified for Raman measurements of the intact muscle providing a reasonable clustering of the lipid containing muscle features present: (1) the dark muscle and the myocommata of the dark muscle area (in the following referred to as Darkfat); (2) the dorsal depot fat and the myocommata of the light muscle at the skin-side of the sample (in the following referred to as Skinfat); (3) the belly flap and the myocommata of the light muscle at the belly-side of the sample (in the following referred to as Infat). The procedure for preparation and spectroscopic measurements was standardised. All samples were randomised, and every day for four successive days 12 or 13 salmon samples were thawed, skinned, measured and homogenised using a onestep homogeniser (Krups). Homogenisation was performed for 20–40 s depending on sample size. All samples were kept cold (0–4 ◦ C) during storage, and after Raman analysis of the ground
87
samples, all samples were stored frozen (−20 ◦ C) in enclosed plastic cans for approximately 1 week. The samples were thawed again prior to oil extraction. Extraction was accomplished by treatment with chloroform and methanol (in a 1:2 volume ratio, respectively) as described in the literature [23]. Forty grams of the homogenised sample were used. At the end of the extraction the chloroform phase was removed by carefully applying reduced pressure and by passing nitrogen gas through the solution (37 ◦ C). The samples were stored cold (0–4 ◦ C) for 1–4 days prior to Raman analysis. 2.4. Raman measurements Raman spectra were collected with a Kaiser Optical Systems Raman RXN1 Analyzer (Kaiser Optical Systems Inc., Ann Arbor, MI, USA) consisting of a holoprobe transmission holographic spectrograph and a charge-coupled device (CCD) detector. The spectrograph was connected with fibre optics to a Kaiser multireaction filtered probehead. The multireaction probe was fitted with a 20 cm long and 12.5 mm o.d. Raman immersion ballprobe (Matrix Solutions, Bothell, WA, USA) incorporating a spherical lens [24]. The Raman system was equipped with a 785nm stabilised external cavity diode laser, and the average laser power was approximately 100 mW at the sample. The detector temperature was −40 ◦ C. Raman spectra of the oils were made with a spectral acquisition time of 5 s, and eight accumulations were averaged for each spectrum. Three replicate spectra were obtained for every oil sample. Raman spectra of the ground salmon samples were made as averages of four accumulations of 20 s each. Five replicate spectra of every ground sample were obtained. Raman spectra of the intact salmon samples were made as averages of four accumulations of 20 s each. Two replicate spectra from the Darkfat area, one replicate spectrum from the Skinfat area and two replicate spectra from the Infat area were obtained. The total acquisition time for each sample matrix were chosen in advance based on the general intensity of the Raman spectra. Restrictions in the number of replicates of intact salmon muscle Raman spectra had to be made due to experimental time limitations. 2.5. Data analysis All Raman spectra were preprocessed in the same way prior to regression analysis. Cosmic spikes were identified and manually removed by comparing replicate spectra of each sample for each data set. A fourth order polynomial was fitted to each Raman spectrum in order to remove the influence of the background fluorescence. The least squares based curve fitting procedure used was based on an iterative fitting process to gradually remove the Raman peaks in order to fit a polynomial directly to the spectrum baseline. The obtained polynomial then was subtracted from the original spectrum [25]. The standard normal variate transformation, which basically is a signal intensity normalisation, was applied on the polynomial-subtracted Raman spectra to eliminate variations in the general intensity levels [26]. This transformation was only applied on the part of the spectrum to be used in regression analysis. After inspection
88
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
of the preprocessed spectra a few of the replicates had to be removed due to abnormal baseline features or spikes deforming Raman bands. Finally, the replicates of every sample of every data set were averaged. The polynomial fitting routine was written in Matlab code (The MathWorks Inc., Natick, MA, USA), whereas all other data analysis were performed using The Unscrambler Software (Camo Process AS, Oslo, Norway). Principal component analysis of the reference GC data revealed that three samples contained a few minor fatty acids (namely phytanic acid, C16:4n − 1 and C24:0) at concentration levels deviating from the other samples. In addition, the reference GC data revealed that the same samples had a higher rate of unidentified peaks compared to the other samples. The samples appeared as clear outliers in the iodine value calibrations of all five data sets. Thus, in order to prevent these samples from severely influencing the iodine value calibrations, the three samples were removed from all data sets prior to regression analysis. In addition, two samples in the oil data set had to be removed due to extensive fluorescence causing saturation of the CCD detector. Preprocessed Raman spectra covering two different frequency regions were used to develop multivariate regression models based on partial least-squares regression (PLSR) [27]. The calculated iodine values were used as reference values, and the optimal number of PLSR factors of the calibration model was determined using full cross-validation [28]. The reference value, yi , and the predicted value, yˆ i , of every sample were used to calculate the prediction error of the cross-validated calibration model, expressed as the root mean square error of cross-validation (RMSECV). The RMSECV value is defined in the following way: N 1 RMSECV = (yi − yˆ i )2 N i=1
where i is the samples from 1 to N. Both the RMSECV and the multivariate correlation coefficient (R) between reference and predicted values were used to evaluate the performance of the regression models. All PLSR models were also validated using a cross model validation routine. Cross model validation is a two-layer crossvalidation, and the method is regarded as more conservative than full cross-validation [29]. It thus may serve as a tool to avoid overfitting in the PLSR models. In the current study 10 segments of randomly chosen samples were used as default in the settings. The cross model validation routine was written in Matlab code.
Fig. 2. Raman spectra of all ground salmon samples. Every spectrum is an average of between three and five replicate spectra. No additional preprocessing have been performed on the spectra.
sets, but also within each set. Raman spectra of all ground salmon samples are shown in Fig. 2, and the figure clearly visualises both intensity variations and significant baseline offsets. In order to obtain robust multivariate calibration models, appropriate preprocessing for removal of these additive and multiplicative effects is crucial. Huge intensity variations between Raman spectra of salmon oil and ground salmon are shown in Fig. 3. Even though the spectral acquisition time for oil was shorter than for ground salmon, the intensities in the oil spectra are considerably higher than in the latter. This difference is due to extensive fluorescence in the oil spectra (as seen by the additive offsets in the figure) and to relative concentration differences. In Raman spectra of biological samples it is often seen that lipid bands are more intense than protein bands. This can be explained by the Raman selectivity rules and the relative concentration of specific Raman scattering bands in fats versus proteins. The ground salmon samples, containing between 5 and 22% fat, are therefore expected to provide lower Raman signals than the oils. Preprocessed Raman spectra of two samples of all five data sets are shown in Fig. 4. The letters in the figure correspond to the Raman band assignments provided in Table 1. As could be expected, the spectra of ground salmon differ the most from the other four types of spectra presented. Apparent differences are seen in the C–H stretch region (regions A and B), and some distinct protein bands only appear in the ground salmon spectra. In addition, close inspection may reveal minor broadening of the bands at 1660 cm−1 (region D) and 1441 cm−1 (region
3. Results and discussion 3.1. Spectral features The intensity of Raman spectra of biological samples depends on a wide range of factors. These include the laser intensity and wavelength, the acquisition time, the concentration of the constituents, and not least the physical properties and texture of the biological matrix. In the present study considerable intensity variations were seen between spectra of the different data
Fig. 3. Two Raman spectra of salmon oil (upper) and two spectra of ground salmon (lower). No preprocessing has been performed on the spectra.
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
89
Table 2 Basic statistics regarding abundant fatty acids and the calculated iodine value
Fig. 4. Preprocessed Raman spectra of two samples of all five data sets. The spectra are separated in vertical direction for clarity. The letters correspond to the Raman band assignments provided in Table 1. Note that the area between 1790 and 2776 cm−1 is not included in the figure.
F) due to protein band inferences. All spectra of oil and ground salmon contained two distinct peaks (at 1521 cm−1 (region E) and 1159 cm−1 (region H)) due to astaxanthin (carotenoids) in the samples. These bands were also present in some of the spectra of the other data sets, and within each data set they provided significant variation. Thus, in all regression analyses carotenoid bands were weighted down to 10% of their original value. In this way the bands were not removed, but their influence on the iodine value predictions were reduced. Two regions were used in the multivariate calibrations, namely the region between 790 and 1790 cm−1 and the region between 2776 and 3052 cm−1 . As seen in Fig. 3, the region between 1790 and 2776 cm−1 carries almost no spectral information and can thereby safely be discarded. The regions below 790 cm−1 often contains useful information, but for the present case these bands are regarded as of minor importance. 3.2. Iodine value predictions One apparent challenge by using bulk fatty acid features to characterise fatty acid profiles is that no bulk parameters alone Table 1 Assignment of important bands in the Raman spectra Regions
Raman shift (cm−1 )
Type of vibration
A
3016
Asymmetric
2935 2900 2854 1749 1660 1521 1441
Asymmetric CH2 stretch Symmetric CH3 stretch Symmetric CH2 stretch C O stretch cis C C stretch/amide I In-phase C C stretch (carotenoids) CH2 scissoring
1302 1266 1159 1122/1081/1066 1004
CH2 twist Symmetric C–H rock Polyene chain C–C stretch (carotenoids) C–C/C–N/C–O stretch Aromatic ring breathing (Phenylalanine)
974 932/866
C–H out-of-plane deformation C–C/C–O stretch
B C D E F G H I J K
The regions are defined in Fig. 3.
C–H stretch
Fatty acid parametersa
Minimum
Maximum
Mean
S.D.b
Uncertainty estimatec
C16:0 C18:1n − 9 C18:2n − 6 C20:1 (sum isomers) C20:5n − 3 (EPA) C22:1 (sum isomers) C22:6n − 3 (DHA) Calculated iodine value
9.4 13.4 2.6 4.3 5.9 3.9 9.8 147.8
13.0 24.9 7.8 10.1 9.6 10.1 13.8 170.0
11.2 20.6 5.9 6.5 7.1 6.2 11.0 159.0
0.9 2.6 1.1 1.4 1 1.4 0.8 5.2
0.17 0.11 0.02 0.07 0.07 0.21 0.19 1.0
a All units of fatty acids provided in percentage of total fat content. All units of iodine value provided in g I2 /100 g fat. b Standard deviation. c The uncertainties are provided as absolute standard deviations.
can describe all the relevant variation in the system. There will always be features not described properly by the parameter chosen, and depending on the system this may interfere with the ability of providing accurate quantitative predictions. The choice of iodine value as the principal parameter of the current study is based on several arguments. First, the iodine value combines information from a large part, though not all, of the fatty acids present. Second, since salmon is a rich source of unsaturated fatty acids the iodine value is regarded as one of many interesting parameters related to quality. Last but not least Raman spectra were expected to provide chemically sound regression coefficients related to the iodine value since Raman bands related to carbon chain unsaturation provide strong Raman signals. Basic statistics regarding the most abundant fatty acids and the calculated iodine value are given in Table 2. The calculated iodine value covers a range of 22.2 units, and the standard deviation clearly indicates that the variation provided by the system is small. Regression results for the iodine values are summarised in Table 3. When comparing both prediction errors and numbers of PLSR factors, the oil spectra expectedly provide better results than the other four data sets. The lowest prediction error (2.5, i.e. 11.3% of the total range) was achieved using a one-component PLSR model. The ground samples provided prediction errors comparable to those of the oils, but here two or three PLSR components were needed. Taking into consideration that the fat concentration of the ground samples are quite low (between 5 and 22%), these are good results. This also shows that even though the overall intensity level of the ground salmon spectra is low, good quantitative information may be obtained. The Infat and Skinfat data sets provided prediction errors comparable but slightly above those of the ground samples, whereas the Darkfat data set provides the highest prediction errors in this study. The regression coefficients of all five data sets are shown in Fig. 5. The oil samples provided the smoothest regression curve, as could be expected for a one-component PLSR model. The other data sets appear to contain higher noise levels, especially in the C–H stretch area. However, for all five data sets the general coefficient pattern is quite clear and similar. There are in particular three dominating peaks present (positioned at 1266, 1660
90
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
Table 3 Summary of regression results for prediction of the iodine value Data type
Number of samples, N
Spectral range (cm−1 )
R
RMSECVa
PLSR factors
Oil
45 45
790–1790 + 2776–3052 790–1790
0.853 0.874
2.7 2.5
1 1
Ground
47 47
790–1790 + 2776–3052 790–1790
0.825 0.855
2.9 2.7
3 2
Darkfat
47 47
790–1790 + 2776–3052 790–1790
0.729 0.666
3.5 3.9
3 4
Skinfat
47 47
790–1790 + 2776–3052 790–1790
0.800 0.762
3.1 3.3
2 2
Infat
47 47
790–1790 + 2776–3052 790–1790
0.827 0.792
2.9 3.1
2 3
a
Units in g I2 /100 g fat.
and 3016 cm−1 , regions G, D and A, respectively). As revealed in Table 1 these bands are all related to the level of unsaturation of carbon–carbon chains, which means that they also should be directly related to the iodine value. The regression coefficients obtained are similar to regression coefficients obtained from earlier model studies for iodine value predictions [18]. The chemical interpretability provided by the regression coefficients is a first indication that the prediction models obtained are robust and not based on spurious correlations. Evaluation of the robustness of the obtained regression models was performed employing a cross model validation routine. For all but three of the PLSR models in Table 3 cross model validation provided prediction errors and model complexities in close proximity to the results obtained by full cross-validation. The two Infat PLSR models as well as one of the Darkfat PLSR models (obtained by including the C–H stretch area) provided substantial increases in the prediction errors, and it is thus likely that slight overfitting has occurred in these PLSR models. A few potential sources of errors related to the current study need mentioning. The uncertainty estimate related to the GC reference measurements is shown in Table 2, but it is also important to remember the uncertainties related to the sampling of reference material. Reference and spectroscopic measurements were performed on the left and the right fillet side of the salmon, respectively, and when dealing with heterogeneous materials differences may appear. For this study a quantitative estimation of these sources of error were however not possible. The extensive
Fig. 5. Regression coefficients of Raman PLSR models for prediction of the iodine value. The number of PLSR components used is given in the parenthesis. The regression coefficients are separated in vertical direction for clarity.
fluorescence observed in some of the salmon oil spectra may be an indication of early lipid oxidation because of extensive treatment of the samples prior to and during oil extraction. This may have influenced the fatty acid composition of the samples to a certain degree. However, fluorescence is a strong phenomenon statistically thousands of times more probable than Raman scattering, hence, early lipid oxidation may significantly affect the fluorescence background of Raman spectra without affecting the actual Raman bands. Last but not least, the spectral preprocessing methods are expected to deal with the majority of the non-chemical variations in the spectra. However, small fluctuations of non-chemical origin will always remain, and in the current study were only small chemical differences are provided, these fluctuations may affect the overall predictive ability of the calibration models. Raman measurements are often based on lasers with spot sizes on the millimetre scale or below, and considerations related to sampling will thus always be crucial when utilising Raman spectroscopy for measurements of heterogeneous biological samples. It has been shown by other authors that small differences in the fatty acid composition of different muscle parts and fat depots of salmon may appear [22]. In the present study the reference measurements were made on extracts of ground salmon muscle, thus there clearly is a mismatch between bulk reference measurements and spectra of intact salmon, and this mismatch may reasonably well explain the difference in calibration results between salmon oil and intact salmon spectra. The same authors [22] also have pointed out that the Darkfat area of the salmon is the region having a fatty acid profile differing the most from the other adipose tissue parts of salmon, and this may serve as a probable explanation why spectra from this specific region clearly provided the highest prediction errors in the present study. However, the results altogether only strengthen the conclusions made from this and other studies that Raman spectroscopy may sensitively reflect small changes in fatty acid profiles. In order to use Raman spectroscopy for determining fatty acid profiles of intact salmon, defined areas of the salmon have to be located, and the location of reference measurements and Raman measurements has to coincide. Apart from rare cases where the variation of a few fatty acids contribute significantly to the overall fatty acid variation, it is
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
questionable if accurate information regarding single fatty acids of animal adipose tissue is obtainable from Raman spectra. The number of different fatty acids is often great, and the difference between chemically closely related fatty acids may be too small to provide significant spectral differences in the Raman spectra, especially when many different fatty acids are present. Secondly, samples from the same animal adipose tissue system often tend to provide the same general fatty acid profile. A consequence is high internal correlation between diverse fatty acids present, and careful steps thus have to be made prior to multivariate calibration in order to thoroughly understand the variation of the system. This was also seen in the present system, and it brought along major limitations in the search for quantifiable fatty acid parameters. The iodine value, for instance, showed a good correlation to the total amount of omega-3 fatty acids present, but there are no clear indications from this or other studies that omega-3 fatty acids may be easily distinguished from other similar fatty acids in Raman spectra of complex fat systems. These kinds of indirect relationships may surely be used directly for quantification of fatty acid features, but with regard to industrial applications the use of indirect relationships may reduce the robustness of the system. An idea for future Raman animal adipose tissue studies is to analyse the covariance patterns related to the Raman spectra and the reference measurements. In this way the latent structures of the data is in focus, not necessarily conventional fatty acid parameters. Thus, the relevant spectral variation of the Raman spectra may be obtained along with the main chemical variation of the adipose tissue system. Attempts of achieving this with the present system were however not successful. Raman spectroscopy clearly has limitations in sensitivity and chemical selectivity being a non-destructive spectroscopic technique. Given the complexity and the small chemical variation provided by the present salmon system, however, the results as revealed in Table 3 are promising. What is in fact seen from the study is that relevant chemical information regarding fatty acid unsaturation may be extracted from all five data sets. In addition, a low number of PLSR components are needed to obtain this information, even for the complex ground salmon samples. The latter point is particularly crucial with respect to obtaining robust multivariate calibrations for possible inprocess Raman applications. Reproducible Raman spectra were obtainable after less than a minute. By comparison, results of GC-analysis are obtainable after hours due to time-consuming sample preparations. In addition to fatty acid analysis, previous studies have shown that Raman spectroscopy also is a potential method for rapid quantification of the content of carotenoids of the salmon muscle [4]. Employing Raman spectroscopy for rapid simultaneous quantification of carotenoids and fatty acid unsaturation of salmon is an additional advantage presenting Raman spectroscopy as a promising technique for in-process salmon analysis. In order to further evaluate the potential of Raman spectroscopy for quantification of fatty acid unsaturation of salmon, several additional samples should be included both for multivariate calibration and prediction. Also means for reducing the effects of the sources of error as listed in this section should be found.
91
4. Conclusions The current study has shown that Raman spectroscopy is a promising method for rapid analysis of the fatty acid unsaturation of salmon. High quality Raman spectra of both intact salmon, ground salmon and salmon oil were easily obtainable, and relevant spectral variation related to fatty acid unsaturation could be extracted from each of the five data sets. Raman spectra of ground salmon provided better iodine value predictions than spectra of intact salmon muscle, and this difference may in part be explained by sampling uncertainties related to Raman measurements of the intact salmon muscle. The results of the ground salmon spectra also show that accurate information regarding fatty acid features is obtainable with Raman spectroscopy even in systems with high contents of protein and water. Iodine value predictions of all five data sets provided PLSR models with few PLSR factors. The potential of Raman spectroscopy as a tool for rapid in-process analysis of fatty acid unsaturation in salmon is demonstrated through high model robustness and low model complexity. Acknowledgements O. Breivik and co-workers at Nutreco (Stavanger, Norway) are acknowledged for providing salmon samples and GC fatty acid reference measurements. G. Enersen at Matforsk is acknowledged for technical assistance, and F. Westad at Matforsk is acknowledged for providing the Matlab code of the cross model validation routine. Financial support from the Agricultural Food Research Foundation is also greatly acknowledged. References [1] Y. Ozaki, R. Cho, K. Ikegaya, S. Muraishi, K. Kawauchi, Appl. Spectrosc. 46 (1992) 1503. [2] E.C.Y. Li-Chan, Trends Food Sci. Technol. 7 (1996) 361. [3] R.J. Beattie, S.J. Bell, L.J. Farmer, B.W. Moss, P.D. Desmond, Meat Sci. 66 (2004) 903. [4] J.P. Wold, B.J. Marquardt, B.K. Dable, D. Robb, B. Hatlen, Appl. Spectrosc. 58 (2004) 395. [5] R. Manoharan, J.J. Baraga, M.S. Feld, R.P. Rava, J. Photochem. Photobiol. B 16 (1992) 211. [6] R. Manoharan, Y. Wang, M.S. Feld, Spectrochim. Acta, Part A 52 (1996) 215. [7] J.F. Brennan, T.J. Romer, R.S. Lees, A.M. Tercyak, J.R. Kramer, M.S. Feld, Circulation 96 (1997) 99. [8] H. Sadeghi-Jorabchi, P.J. Hendra, R.H. Wilson, P.S. Belton, J. Am. Oil Chem. Soc. 67 (1990) 483. [9] R.C. Barthus, R.J. Poppi, Vib. Spectrosc. 26 (2001) 99. [10] G.F. Bailey, R.J. Horvat, J. Am. Oil Chem. Soc. 49 (1972) 494. [11] H. Sadeghi-Jorabchi, R.H. Wilson, P.S. Belton, J.D. Edwardswebb, D.T. Coxon, Spectrochim. Acta, Part A 47 (1991) 1449. [12] M. Meurens, V. Baeten, S.H. Yan, E. Mignolet, Y. Larondelle, J. Agric. Food Chem. 53 (2005) 5831. [13] V. Baeten, P. Hourant, M.T. Morales, R. Aparicio, J. Agric. Food Chem. 46 (1998) 2638. [14] J.R. Beattie, S.E.J. Bell, C. Borgaard, A.M. Fearon, B.W. Moss, Lipids 39 (2004) 897. [15] C. Paquot, A. Hautfenne, IUPAC Standard Methods for the Analysis of Oils, Fats and Derivatives, Blackwell, Oxford, 1987.
92
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
[16] H. Li, F.R. van de Voort, J. Sedman, A.A. Ismail, J. Am. Oil Chem. Soc. 76 (1999) 491. [17] J. Sedman, F.R. van de Voort, A.A. Ismail, P. Maes, J. Am. Oil Chem. Soc. 75 (1998) 33. [18] N.K. Afseth, V.H. Segtnan, B.J. Marquardt, J.P. Wold, Appl. Spectrosc. 59 (2005) 1324. [19] Norwegian Standard 9401 Atlantic Salmon: Reference Sampling for Quality Judgement, 1994, p. 3 (translation from Norwegian). [20] O. Grahl-Nielsen, T. Barnung, Mar. Environ. Res. 17 (1985) 218. [21] R.E. Walpole, R.H. Myers, S.L. Myers, Probability and Statistics for Engineers and Scientists, Prentice Hall International, New Jersey, 1998.
[22] M. Aursand, B. Bleivik, J.R. Rainuzzo, L. Jorgensen, V. Mohr, J. Sci. Food Agric. 64 (1994) 239. [23] E.G. Bligh, W.J. Dyer, Can. J. Biochem. Physiol. 37 (1959) 911. [24] B.J. Marquardt, T. Le, L.W. Burgess, Proc. SPIE-Int. Soc. Opt. Eng. 4469 (2001) 62. [25] C.A. Lieber, A. Mahadevan-Jansen, Appl. Spectrosc. 57 (2003) 1363. [26] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Appl. Spectrosc. 43 (1989) 772. [27] H. Martens, T. Næs, Multivariate Calibration, John Wiley & Sons, Chichester, 1989. [28] M. Stone, J. Roy. Stat. Soc. Ser. B: Methodol. 36 (1974) 111. [29] N.J. Le Roux, S.J. Steel, N. Louw, J. Stat. Comput. Simul. 59 (1997) 195.
The potential of Raman spectroscopy for characterisation of the fatty acid unsaturation of salmon Nils Kristian Afseth a,b,∗ , Jens Petter Wold a , Vegard Herman Segtnan a a
b
˚ Norway MATFORSK, Norwegian Food Research Institute, Osloveien 1, N-1430 As, ˚ Norway Department of Chemistry, Biotechnology and Food Science, Norwegian University of Life Sciences, As, Received 8 February 2006; received in revised form 3 May 2006; accepted 4 May 2006 Available online 10 May 2006
Abstract Raman spectroscopy has been evaluated for characterisation of the degree of fatty acid unsaturation (iodine value) of salmon (Salmo salar). The Norwegian Quality Cuts from 50 salmon samples were obtained, and the samples provided an iodine value range of 147.8–170.0 g I2 /100 g fat, reflecting a normal variation of farmed salmon. Raman measurements were performed both on different spots of the intact salmon muscle, on ground salmon samples as well as on oil extracts, and partial least squares regression (PLSR) was utilised for calibration. The oil spectra provided better iodine value predictions than the other data sets, and a correlation coefficient of 0.87 with a root mean square error of cross-validation of 2.5 g I2 /100 g fat was achieved using only one PLSR component. The ground samples provided comparable results, but at least two PLSR components were needed. Higher prediction errors were obtained from Raman spectra of intact salmon muscle, and this may partly be explained by sampling uncertainties in the relation between Raman measurements and reference analysis. All PLSR models obtained were based on chemically sound regression coefficients, and thus information regarding fatty acid unsaturation is readily available from Raman spectra even in systems with high contents of protein and water. The accuracy, the robustness and the low complexity of the PLSR models obtained suggest Raman spectroscopy as a promising method for rapid in-process control of the degree of unsaturation in salmon samples. © 2006 Elsevier B.V. All rights reserved. Keywords: Raman spectroscopy; Salmon; Iodine value; Partial least squares regression; Sampling; In-process control
1. Introduction There is an increasing interest in potential methods for fast and non-destructive characterisation of the fatty acid composition of marine food products. This is related to the great commercial value connected to dietary fats of marine origin in general, and the presence of several essential health promoting fatty acids like eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA) in particular. If reliable in-process techniques for control of fatty acid features are implemented, this will benefit both producers and consumers. For the producers, in-process control may provide better means for sorting raw materials as well as provide better documentation of product quality for price differentiation and promotion. In addition, rapid control of fatty acid composition is essential when controlling the effect of different feeding regimes. Ever increasing consumer demands for
∗
Corresponding author. Tel.: +47 64 97 04 18; fax: +47 64 97 03 33. E-mail address: [email protected] (N.K. Afseth).
0003-2670/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2006.05.013
improved food component labelling will also be met. At present, gas chromatography (GC) is the method of choice for analysing the fatty acid composition of foods. This method, however, is destructive and involves laborious wet chemistry for sample preparation. The results obtained are accurate, but the procedure is often too costly and time consuming for industrial quality evaluation. For several years Raman spectroscopy has been referred to as a promising method for rapid and non-destructive analysis of foods [1,2]. The technique is based on measuring fundamental vibrational modes of chemical bonds, and quantitative as well as qualitative structural and chemical information of a given sample may be obtained. What makes Raman spectroscopy especially suitable for food analysis is that minimal or no sample preparation is required, and that measurements may be carried out utilising fibre optics. In addition, the influence of water on the Raman spectra is almost negligible. Interference from strong and broad fluorescence signals combined with low sensitivity have traditionally been some of the major challenges. However, instrumental improvements concerning
86
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
both sampling techniques, stability and sensitivity have been achieved, paving the way for an increasing number of articles related to Raman spectroscopy and food analysis in recent years [3,4]. A range of feasibility studies has shown the potential of employing Raman spectroscopy as a means for quantifying different parameters related to fat composition. In medical diagnostics, quantification of different lipid groups like phospholipids, cholesterol, cholesterolesters and various fatty acids in arteries, skin layers or other kinds of healthy and diseased tissues has been accomplished [5–7]. In food science the focus has mainly been on the analysis of pure oils and fats. First and foremost, bulk parameters like the iodine value have been shown to correlate well with Raman spectra [8,9]. This is mainly related to the favourable Raman cross sections of chemical bands originating from the unsaturation of carbon chains. Good quantitative descriptions of the cis/trans-isomer ratio of fats and oils [10,11] as well as the level of conjugated fatty acids in milk fat [12] have also been obtained. In addition, oil and fat classification [13] and predictions of various single fatty acids have been carried out [14]. However, there are few studies published focusing on Raman analysis of fat composition of fish and animal adipose tissue, and to the knowledge of the authors there are no recorded references utilising Raman spectroscopy for quantifying fat composition of food systems with high contents of water and proteins. Traditionally, the iodine value has been one of the standard parameters related to chemical composition and quality of fats and oils. The analytical test is based on the ability of elemental iodine (I2 ) to quantitatively add to carbon–carbon double bonds, and the parameter is defined as the weight of iodine absorbed in an amount of 100 g of fat sample. Thus, as long as there are no other compounds present containing carbon–carbon double bonds, the iodine value is a direct measure of the level of fatty acid unsaturation [15]. Spectroscopic techniques like near-infrared (NIR) and mid-infrared (IR) spectroscopy both have been used for rapid determination of the iodine value of fats and oils with high accuracy [16,17]. However, strong water absorbances severely influence both NIR and IR spectra. In addition, protein absorbances often may obscure NIR bands related to fatty acid unsaturation [18]. Thus, these techniques have major disadvantages when analysing samples with high contents of water and proteins. Similar disturbances may to a great extent be avoided using Raman spectroscopy. In a previous study, the authors showed that Raman spectroscopy may be used to quantify bulk fatty acid features like the iodine value and the amount of saturated, monounsaturated and polyunsaturated fatty acids of complex food model samples with high contents of proteins and water [18]. The samples had a wide range in the variation of the fatty acid composition, and important spectral characteristics regarding the fatty acid parameters were obtained. In the present study a set of salmon samples with naturally varying fatty acid composition and overall sample heterogeneity has been analysed. The main objectives of the study were two-fold. First, the performance of Raman spectroscopy for modelling and prediction of the iodine
values of salmon was investigated. Second, by retrieving spectra from different parts of the intact salmon muscle, ground samples and oil extracts, a comparison of different sampling regimes for spectral acquisition was conducted. 2. Experimental 2.1. Materials Fifty samples of Atlantic salmon (S. salar) were obtained from the Nutreco Aquaculture Research Centre (Stavanger, Norway). The salmon samples originated from 13 different farms, and they were collected during a period of approximately 2 months. The samples were selected in order to span the range of the fatty acid composition normally occurring in farmed salmon. Muscle samples from the region of the Norwegian Quality Cut (NQC) were obtained [19], and the NQCs were split in half. One half of the NQC was left for reference analysis at Nutreco, whereas the other half was shipped frozen in enclosed plastic ˚ Norway) for Raman spectroscopic analbags to Matforsk (As, ysis. The location and the splitting of the NQCs are indicated in Fig. 1. 2.2. Reference analysis The fatty acid composition of the salmon samples were determined using a standard gas chromatography (GC) procedure. Prior to sample homogenisation the skin of the NQC was removed such that neither skin remnants were left at the muscle nor muscle remnants were left at the skin. Fat extraction
Fig. 1. Location and anatomy of the Norwegian Quality Cut (NQC) of salmon. The definitions used in the current study are provided in parenthesis. The dotted square indicates the division of NQCs into samples for reference and spectroscopic measurements, respectively.
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
and conversion to methyl esters were performed as described in the literature [20]. Methanolic hydrochloric acid (2.15 mL, 3N) was added to a sample of homogenised salmon (200 mg) and left enclosed over night. Water (1 mL) was added, and the methyl esters were extracted from the methanol phase with hexane (2 mL). The sample was centrifuged twice, with collection of the hexane phase and re-addition of hexane (2 mL) in between. The final hexane phase (25 L) was diluted with isooctane (1 mL) and injected on a Perkin-Elmer Autosystem GC equipped with a PSS injector, a CP-Wax 52 CB column (L = 25 m; i.d. = 0.25 mm; d.f. = 0.20 m, Varian), a flame ionisation detector and He as carrier gas. Extraction and GC-analysis were performed in duplicates of all samples. The compounds were identified and system performance was checked using standard samples. The concentration of individual fatty acids was expressed in percentage of total fat content. The iodine value was calculated directly from the GC results by converting the concentration of 38 fatty acids (expressed in percentage of total fat content) to the corresponding number of carbon–carbon double bonds (expressed in mol). The conversion was effected employing the molar mass of each corresponding fatty acid methyl ester. The iodine value then was calculated as the weight of elemental iodine corresponding to the total number of carbon–carbon double bonds. The uncertainty of the calculated iodine value was estimated using the Gaussian law of error propagation [21], whereas uncertainty estimates of each individual fatty acid were calculated on the basis of replicate measurements of a routine fatty acid standard. 2.3. Sample preparation A total of five data sets of Raman spectra were obtained from the salmon samples: one data set from extracted oils; one data set from ground salmon samples and three different data sets from intact samples. The main parts of the salmon NQCs are shown in Fig. 1. Lipids are mainly located in the dorsal fat depot, in the belly flap, in the collagenous connective tissue (myocommata) throughout the light and dark muscles, as well as in the intermediate space of myofibrils in the dark muscle [22]. The salmon samples showed great variability in both size and structure. Thus, three main regions were identified for Raman measurements of the intact muscle providing a reasonable clustering of the lipid containing muscle features present: (1) the dark muscle and the myocommata of the dark muscle area (in the following referred to as Darkfat); (2) the dorsal depot fat and the myocommata of the light muscle at the skin-side of the sample (in the following referred to as Skinfat); (3) the belly flap and the myocommata of the light muscle at the belly-side of the sample (in the following referred to as Infat). The procedure for preparation and spectroscopic measurements was standardised. All samples were randomised, and every day for four successive days 12 or 13 salmon samples were thawed, skinned, measured and homogenised using a onestep homogeniser (Krups). Homogenisation was performed for 20–40 s depending on sample size. All samples were kept cold (0–4 ◦ C) during storage, and after Raman analysis of the ground
87
samples, all samples were stored frozen (−20 ◦ C) in enclosed plastic cans for approximately 1 week. The samples were thawed again prior to oil extraction. Extraction was accomplished by treatment with chloroform and methanol (in a 1:2 volume ratio, respectively) as described in the literature [23]. Forty grams of the homogenised sample were used. At the end of the extraction the chloroform phase was removed by carefully applying reduced pressure and by passing nitrogen gas through the solution (37 ◦ C). The samples were stored cold (0–4 ◦ C) for 1–4 days prior to Raman analysis. 2.4. Raman measurements Raman spectra were collected with a Kaiser Optical Systems Raman RXN1 Analyzer (Kaiser Optical Systems Inc., Ann Arbor, MI, USA) consisting of a holoprobe transmission holographic spectrograph and a charge-coupled device (CCD) detector. The spectrograph was connected with fibre optics to a Kaiser multireaction filtered probehead. The multireaction probe was fitted with a 20 cm long and 12.5 mm o.d. Raman immersion ballprobe (Matrix Solutions, Bothell, WA, USA) incorporating a spherical lens [24]. The Raman system was equipped with a 785nm stabilised external cavity diode laser, and the average laser power was approximately 100 mW at the sample. The detector temperature was −40 ◦ C. Raman spectra of the oils were made with a spectral acquisition time of 5 s, and eight accumulations were averaged for each spectrum. Three replicate spectra were obtained for every oil sample. Raman spectra of the ground salmon samples were made as averages of four accumulations of 20 s each. Five replicate spectra of every ground sample were obtained. Raman spectra of the intact salmon samples were made as averages of four accumulations of 20 s each. Two replicate spectra from the Darkfat area, one replicate spectrum from the Skinfat area and two replicate spectra from the Infat area were obtained. The total acquisition time for each sample matrix were chosen in advance based on the general intensity of the Raman spectra. Restrictions in the number of replicates of intact salmon muscle Raman spectra had to be made due to experimental time limitations. 2.5. Data analysis All Raman spectra were preprocessed in the same way prior to regression analysis. Cosmic spikes were identified and manually removed by comparing replicate spectra of each sample for each data set. A fourth order polynomial was fitted to each Raman spectrum in order to remove the influence of the background fluorescence. The least squares based curve fitting procedure used was based on an iterative fitting process to gradually remove the Raman peaks in order to fit a polynomial directly to the spectrum baseline. The obtained polynomial then was subtracted from the original spectrum [25]. The standard normal variate transformation, which basically is a signal intensity normalisation, was applied on the polynomial-subtracted Raman spectra to eliminate variations in the general intensity levels [26]. This transformation was only applied on the part of the spectrum to be used in regression analysis. After inspection
88
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
of the preprocessed spectra a few of the replicates had to be removed due to abnormal baseline features or spikes deforming Raman bands. Finally, the replicates of every sample of every data set were averaged. The polynomial fitting routine was written in Matlab code (The MathWorks Inc., Natick, MA, USA), whereas all other data analysis were performed using The Unscrambler Software (Camo Process AS, Oslo, Norway). Principal component analysis of the reference GC data revealed that three samples contained a few minor fatty acids (namely phytanic acid, C16:4n − 1 and C24:0) at concentration levels deviating from the other samples. In addition, the reference GC data revealed that the same samples had a higher rate of unidentified peaks compared to the other samples. The samples appeared as clear outliers in the iodine value calibrations of all five data sets. Thus, in order to prevent these samples from severely influencing the iodine value calibrations, the three samples were removed from all data sets prior to regression analysis. In addition, two samples in the oil data set had to be removed due to extensive fluorescence causing saturation of the CCD detector. Preprocessed Raman spectra covering two different frequency regions were used to develop multivariate regression models based on partial least-squares regression (PLSR) [27]. The calculated iodine values were used as reference values, and the optimal number of PLSR factors of the calibration model was determined using full cross-validation [28]. The reference value, yi , and the predicted value, yˆ i , of every sample were used to calculate the prediction error of the cross-validated calibration model, expressed as the root mean square error of cross-validation (RMSECV). The RMSECV value is defined in the following way: N 1 RMSECV = (yi − yˆ i )2 N i=1
where i is the samples from 1 to N. Both the RMSECV and the multivariate correlation coefficient (R) between reference and predicted values were used to evaluate the performance of the regression models. All PLSR models were also validated using a cross model validation routine. Cross model validation is a two-layer crossvalidation, and the method is regarded as more conservative than full cross-validation [29]. It thus may serve as a tool to avoid overfitting in the PLSR models. In the current study 10 segments of randomly chosen samples were used as default in the settings. The cross model validation routine was written in Matlab code.
Fig. 2. Raman spectra of all ground salmon samples. Every spectrum is an average of between three and five replicate spectra. No additional preprocessing have been performed on the spectra.
sets, but also within each set. Raman spectra of all ground salmon samples are shown in Fig. 2, and the figure clearly visualises both intensity variations and significant baseline offsets. In order to obtain robust multivariate calibration models, appropriate preprocessing for removal of these additive and multiplicative effects is crucial. Huge intensity variations between Raman spectra of salmon oil and ground salmon are shown in Fig. 3. Even though the spectral acquisition time for oil was shorter than for ground salmon, the intensities in the oil spectra are considerably higher than in the latter. This difference is due to extensive fluorescence in the oil spectra (as seen by the additive offsets in the figure) and to relative concentration differences. In Raman spectra of biological samples it is often seen that lipid bands are more intense than protein bands. This can be explained by the Raman selectivity rules and the relative concentration of specific Raman scattering bands in fats versus proteins. The ground salmon samples, containing between 5 and 22% fat, are therefore expected to provide lower Raman signals than the oils. Preprocessed Raman spectra of two samples of all five data sets are shown in Fig. 4. The letters in the figure correspond to the Raman band assignments provided in Table 1. As could be expected, the spectra of ground salmon differ the most from the other four types of spectra presented. Apparent differences are seen in the C–H stretch region (regions A and B), and some distinct protein bands only appear in the ground salmon spectra. In addition, close inspection may reveal minor broadening of the bands at 1660 cm−1 (region D) and 1441 cm−1 (region
3. Results and discussion 3.1. Spectral features The intensity of Raman spectra of biological samples depends on a wide range of factors. These include the laser intensity and wavelength, the acquisition time, the concentration of the constituents, and not least the physical properties and texture of the biological matrix. In the present study considerable intensity variations were seen between spectra of the different data
Fig. 3. Two Raman spectra of salmon oil (upper) and two spectra of ground salmon (lower). No preprocessing has been performed on the spectra.
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
89
Table 2 Basic statistics regarding abundant fatty acids and the calculated iodine value
Fig. 4. Preprocessed Raman spectra of two samples of all five data sets. The spectra are separated in vertical direction for clarity. The letters correspond to the Raman band assignments provided in Table 1. Note that the area between 1790 and 2776 cm−1 is not included in the figure.
F) due to protein band inferences. All spectra of oil and ground salmon contained two distinct peaks (at 1521 cm−1 (region E) and 1159 cm−1 (region H)) due to astaxanthin (carotenoids) in the samples. These bands were also present in some of the spectra of the other data sets, and within each data set they provided significant variation. Thus, in all regression analyses carotenoid bands were weighted down to 10% of their original value. In this way the bands were not removed, but their influence on the iodine value predictions were reduced. Two regions were used in the multivariate calibrations, namely the region between 790 and 1790 cm−1 and the region between 2776 and 3052 cm−1 . As seen in Fig. 3, the region between 1790 and 2776 cm−1 carries almost no spectral information and can thereby safely be discarded. The regions below 790 cm−1 often contains useful information, but for the present case these bands are regarded as of minor importance. 3.2. Iodine value predictions One apparent challenge by using bulk fatty acid features to characterise fatty acid profiles is that no bulk parameters alone Table 1 Assignment of important bands in the Raman spectra Regions
Raman shift (cm−1 )
Type of vibration
A
3016
Asymmetric
2935 2900 2854 1749 1660 1521 1441
Asymmetric CH2 stretch Symmetric CH3 stretch Symmetric CH2 stretch C O stretch cis C C stretch/amide I In-phase C C stretch (carotenoids) CH2 scissoring
1302 1266 1159 1122/1081/1066 1004
CH2 twist Symmetric C–H rock Polyene chain C–C stretch (carotenoids) C–C/C–N/C–O stretch Aromatic ring breathing (Phenylalanine)
974 932/866
C–H out-of-plane deformation C–C/C–O stretch
B C D E F G H I J K
The regions are defined in Fig. 3.
C–H stretch
Fatty acid parametersa
Minimum
Maximum
Mean
S.D.b
Uncertainty estimatec
C16:0 C18:1n − 9 C18:2n − 6 C20:1 (sum isomers) C20:5n − 3 (EPA) C22:1 (sum isomers) C22:6n − 3 (DHA) Calculated iodine value
9.4 13.4 2.6 4.3 5.9 3.9 9.8 147.8
13.0 24.9 7.8 10.1 9.6 10.1 13.8 170.0
11.2 20.6 5.9 6.5 7.1 6.2 11.0 159.0
0.9 2.6 1.1 1.4 1 1.4 0.8 5.2
0.17 0.11 0.02 0.07 0.07 0.21 0.19 1.0
a All units of fatty acids provided in percentage of total fat content. All units of iodine value provided in g I2 /100 g fat. b Standard deviation. c The uncertainties are provided as absolute standard deviations.
can describe all the relevant variation in the system. There will always be features not described properly by the parameter chosen, and depending on the system this may interfere with the ability of providing accurate quantitative predictions. The choice of iodine value as the principal parameter of the current study is based on several arguments. First, the iodine value combines information from a large part, though not all, of the fatty acids present. Second, since salmon is a rich source of unsaturated fatty acids the iodine value is regarded as one of many interesting parameters related to quality. Last but not least Raman spectra were expected to provide chemically sound regression coefficients related to the iodine value since Raman bands related to carbon chain unsaturation provide strong Raman signals. Basic statistics regarding the most abundant fatty acids and the calculated iodine value are given in Table 2. The calculated iodine value covers a range of 22.2 units, and the standard deviation clearly indicates that the variation provided by the system is small. Regression results for the iodine values are summarised in Table 3. When comparing both prediction errors and numbers of PLSR factors, the oil spectra expectedly provide better results than the other four data sets. The lowest prediction error (2.5, i.e. 11.3% of the total range) was achieved using a one-component PLSR model. The ground samples provided prediction errors comparable to those of the oils, but here two or three PLSR components were needed. Taking into consideration that the fat concentration of the ground samples are quite low (between 5 and 22%), these are good results. This also shows that even though the overall intensity level of the ground salmon spectra is low, good quantitative information may be obtained. The Infat and Skinfat data sets provided prediction errors comparable but slightly above those of the ground samples, whereas the Darkfat data set provides the highest prediction errors in this study. The regression coefficients of all five data sets are shown in Fig. 5. The oil samples provided the smoothest regression curve, as could be expected for a one-component PLSR model. The other data sets appear to contain higher noise levels, especially in the C–H stretch area. However, for all five data sets the general coefficient pattern is quite clear and similar. There are in particular three dominating peaks present (positioned at 1266, 1660
90
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
Table 3 Summary of regression results for prediction of the iodine value Data type
Number of samples, N
Spectral range (cm−1 )
R
RMSECVa
PLSR factors
Oil
45 45
790–1790 + 2776–3052 790–1790
0.853 0.874
2.7 2.5
1 1
Ground
47 47
790–1790 + 2776–3052 790–1790
0.825 0.855
2.9 2.7
3 2
Darkfat
47 47
790–1790 + 2776–3052 790–1790
0.729 0.666
3.5 3.9
3 4
Skinfat
47 47
790–1790 + 2776–3052 790–1790
0.800 0.762
3.1 3.3
2 2
Infat
47 47
790–1790 + 2776–3052 790–1790
0.827 0.792
2.9 3.1
2 3
a
Units in g I2 /100 g fat.
and 3016 cm−1 , regions G, D and A, respectively). As revealed in Table 1 these bands are all related to the level of unsaturation of carbon–carbon chains, which means that they also should be directly related to the iodine value. The regression coefficients obtained are similar to regression coefficients obtained from earlier model studies for iodine value predictions [18]. The chemical interpretability provided by the regression coefficients is a first indication that the prediction models obtained are robust and not based on spurious correlations. Evaluation of the robustness of the obtained regression models was performed employing a cross model validation routine. For all but three of the PLSR models in Table 3 cross model validation provided prediction errors and model complexities in close proximity to the results obtained by full cross-validation. The two Infat PLSR models as well as one of the Darkfat PLSR models (obtained by including the C–H stretch area) provided substantial increases in the prediction errors, and it is thus likely that slight overfitting has occurred in these PLSR models. A few potential sources of errors related to the current study need mentioning. The uncertainty estimate related to the GC reference measurements is shown in Table 2, but it is also important to remember the uncertainties related to the sampling of reference material. Reference and spectroscopic measurements were performed on the left and the right fillet side of the salmon, respectively, and when dealing with heterogeneous materials differences may appear. For this study a quantitative estimation of these sources of error were however not possible. The extensive
Fig. 5. Regression coefficients of Raman PLSR models for prediction of the iodine value. The number of PLSR components used is given in the parenthesis. The regression coefficients are separated in vertical direction for clarity.
fluorescence observed in some of the salmon oil spectra may be an indication of early lipid oxidation because of extensive treatment of the samples prior to and during oil extraction. This may have influenced the fatty acid composition of the samples to a certain degree. However, fluorescence is a strong phenomenon statistically thousands of times more probable than Raman scattering, hence, early lipid oxidation may significantly affect the fluorescence background of Raman spectra without affecting the actual Raman bands. Last but not least, the spectral preprocessing methods are expected to deal with the majority of the non-chemical variations in the spectra. However, small fluctuations of non-chemical origin will always remain, and in the current study were only small chemical differences are provided, these fluctuations may affect the overall predictive ability of the calibration models. Raman measurements are often based on lasers with spot sizes on the millimetre scale or below, and considerations related to sampling will thus always be crucial when utilising Raman spectroscopy for measurements of heterogeneous biological samples. It has been shown by other authors that small differences in the fatty acid composition of different muscle parts and fat depots of salmon may appear [22]. In the present study the reference measurements were made on extracts of ground salmon muscle, thus there clearly is a mismatch between bulk reference measurements and spectra of intact salmon, and this mismatch may reasonably well explain the difference in calibration results between salmon oil and intact salmon spectra. The same authors [22] also have pointed out that the Darkfat area of the salmon is the region having a fatty acid profile differing the most from the other adipose tissue parts of salmon, and this may serve as a probable explanation why spectra from this specific region clearly provided the highest prediction errors in the present study. However, the results altogether only strengthen the conclusions made from this and other studies that Raman spectroscopy may sensitively reflect small changes in fatty acid profiles. In order to use Raman spectroscopy for determining fatty acid profiles of intact salmon, defined areas of the salmon have to be located, and the location of reference measurements and Raman measurements has to coincide. Apart from rare cases where the variation of a few fatty acids contribute significantly to the overall fatty acid variation, it is
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
questionable if accurate information regarding single fatty acids of animal adipose tissue is obtainable from Raman spectra. The number of different fatty acids is often great, and the difference between chemically closely related fatty acids may be too small to provide significant spectral differences in the Raman spectra, especially when many different fatty acids are present. Secondly, samples from the same animal adipose tissue system often tend to provide the same general fatty acid profile. A consequence is high internal correlation between diverse fatty acids present, and careful steps thus have to be made prior to multivariate calibration in order to thoroughly understand the variation of the system. This was also seen in the present system, and it brought along major limitations in the search for quantifiable fatty acid parameters. The iodine value, for instance, showed a good correlation to the total amount of omega-3 fatty acids present, but there are no clear indications from this or other studies that omega-3 fatty acids may be easily distinguished from other similar fatty acids in Raman spectra of complex fat systems. These kinds of indirect relationships may surely be used directly for quantification of fatty acid features, but with regard to industrial applications the use of indirect relationships may reduce the robustness of the system. An idea for future Raman animal adipose tissue studies is to analyse the covariance patterns related to the Raman spectra and the reference measurements. In this way the latent structures of the data is in focus, not necessarily conventional fatty acid parameters. Thus, the relevant spectral variation of the Raman spectra may be obtained along with the main chemical variation of the adipose tissue system. Attempts of achieving this with the present system were however not successful. Raman spectroscopy clearly has limitations in sensitivity and chemical selectivity being a non-destructive spectroscopic technique. Given the complexity and the small chemical variation provided by the present salmon system, however, the results as revealed in Table 3 are promising. What is in fact seen from the study is that relevant chemical information regarding fatty acid unsaturation may be extracted from all five data sets. In addition, a low number of PLSR components are needed to obtain this information, even for the complex ground salmon samples. The latter point is particularly crucial with respect to obtaining robust multivariate calibrations for possible inprocess Raman applications. Reproducible Raman spectra were obtainable after less than a minute. By comparison, results of GC-analysis are obtainable after hours due to time-consuming sample preparations. In addition to fatty acid analysis, previous studies have shown that Raman spectroscopy also is a potential method for rapid quantification of the content of carotenoids of the salmon muscle [4]. Employing Raman spectroscopy for rapid simultaneous quantification of carotenoids and fatty acid unsaturation of salmon is an additional advantage presenting Raman spectroscopy as a promising technique for in-process salmon analysis. In order to further evaluate the potential of Raman spectroscopy for quantification of fatty acid unsaturation of salmon, several additional samples should be included both for multivariate calibration and prediction. Also means for reducing the effects of the sources of error as listed in this section should be found.
91
4. Conclusions The current study has shown that Raman spectroscopy is a promising method for rapid analysis of the fatty acid unsaturation of salmon. High quality Raman spectra of both intact salmon, ground salmon and salmon oil were easily obtainable, and relevant spectral variation related to fatty acid unsaturation could be extracted from each of the five data sets. Raman spectra of ground salmon provided better iodine value predictions than spectra of intact salmon muscle, and this difference may in part be explained by sampling uncertainties related to Raman measurements of the intact salmon muscle. The results of the ground salmon spectra also show that accurate information regarding fatty acid features is obtainable with Raman spectroscopy even in systems with high contents of protein and water. Iodine value predictions of all five data sets provided PLSR models with few PLSR factors. The potential of Raman spectroscopy as a tool for rapid in-process analysis of fatty acid unsaturation in salmon is demonstrated through high model robustness and low model complexity. Acknowledgements O. Breivik and co-workers at Nutreco (Stavanger, Norway) are acknowledged for providing salmon samples and GC fatty acid reference measurements. G. Enersen at Matforsk is acknowledged for technical assistance, and F. Westad at Matforsk is acknowledged for providing the Matlab code of the cross model validation routine. Financial support from the Agricultural Food Research Foundation is also greatly acknowledged. References [1] Y. Ozaki, R. Cho, K. Ikegaya, S. Muraishi, K. Kawauchi, Appl. Spectrosc. 46 (1992) 1503. [2] E.C.Y. Li-Chan, Trends Food Sci. Technol. 7 (1996) 361. [3] R.J. Beattie, S.J. Bell, L.J. Farmer, B.W. Moss, P.D. Desmond, Meat Sci. 66 (2004) 903. [4] J.P. Wold, B.J. Marquardt, B.K. Dable, D. Robb, B. Hatlen, Appl. Spectrosc. 58 (2004) 395. [5] R. Manoharan, J.J. Baraga, M.S. Feld, R.P. Rava, J. Photochem. Photobiol. B 16 (1992) 211. [6] R. Manoharan, Y. Wang, M.S. Feld, Spectrochim. Acta, Part A 52 (1996) 215. [7] J.F. Brennan, T.J. Romer, R.S. Lees, A.M. Tercyak, J.R. Kramer, M.S. Feld, Circulation 96 (1997) 99. [8] H. Sadeghi-Jorabchi, P.J. Hendra, R.H. Wilson, P.S. Belton, J. Am. Oil Chem. Soc. 67 (1990) 483. [9] R.C. Barthus, R.J. Poppi, Vib. Spectrosc. 26 (2001) 99. [10] G.F. Bailey, R.J. Horvat, J. Am. Oil Chem. Soc. 49 (1972) 494. [11] H. Sadeghi-Jorabchi, R.H. Wilson, P.S. Belton, J.D. Edwardswebb, D.T. Coxon, Spectrochim. Acta, Part A 47 (1991) 1449. [12] M. Meurens, V. Baeten, S.H. Yan, E. Mignolet, Y. Larondelle, J. Agric. Food Chem. 53 (2005) 5831. [13] V. Baeten, P. Hourant, M.T. Morales, R. Aparicio, J. Agric. Food Chem. 46 (1998) 2638. [14] J.R. Beattie, S.E.J. Bell, C. Borgaard, A.M. Fearon, B.W. Moss, Lipids 39 (2004) 897. [15] C. Paquot, A. Hautfenne, IUPAC Standard Methods for the Analysis of Oils, Fats and Derivatives, Blackwell, Oxford, 1987.
92
N.K. Afseth et al. / Analytica Chimica Acta 572 (2006) 85–92
[16] H. Li, F.R. van de Voort, J. Sedman, A.A. Ismail, J. Am. Oil Chem. Soc. 76 (1999) 491. [17] J. Sedman, F.R. van de Voort, A.A. Ismail, P. Maes, J. Am. Oil Chem. Soc. 75 (1998) 33. [18] N.K. Afseth, V.H. Segtnan, B.J. Marquardt, J.P. Wold, Appl. Spectrosc. 59 (2005) 1324. [19] Norwegian Standard 9401 Atlantic Salmon: Reference Sampling for Quality Judgement, 1994, p. 3 (translation from Norwegian). [20] O. Grahl-Nielsen, T. Barnung, Mar. Environ. Res. 17 (1985) 218. [21] R.E. Walpole, R.H. Myers, S.L. Myers, Probability and Statistics for Engineers and Scientists, Prentice Hall International, New Jersey, 1998.
[22] M. Aursand, B. Bleivik, J.R. Rainuzzo, L. Jorgensen, V. Mohr, J. Sci. Food Agric. 64 (1994) 239. [23] E.G. Bligh, W.J. Dyer, Can. J. Biochem. Physiol. 37 (1959) 911. [24] B.J. Marquardt, T. Le, L.W. Burgess, Proc. SPIE-Int. Soc. Opt. Eng. 4469 (2001) 62. [25] C.A. Lieber, A. Mahadevan-Jansen, Appl. Spectrosc. 57 (2003) 1363. [26] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Appl. Spectrosc. 43 (1989) 772. [27] H. Martens, T. Næs, Multivariate Calibration, John Wiley & Sons, Chichester, 1989. [28] M. Stone, J. Roy. Stat. Soc. Ser. B: Methodol. 36 (1974) 111. [29] N.J. Le Roux, S.J. Steel, N. Louw, J. Stat. Comput. Simul. 59 (1997) 195.